Generating lay summaries of biomedical research remains a time-intensive task, despite their importance in bridging the gap between scientific findings and non-expert audiences. This study introduces a retrieval-augmented fine-tuning framework for biomedical lay summarization, integrating abstract-driven semantic retrieval with LoRA-tuned LLaMA 3.1 models. Abstracts are used as queries to retrieve relevant text segments from full-text articles, which are then incorporated into prompts for supervised fine-tuning. Evaluations on the PLOS and eLife datasets show that this hybrid approach significantly improves relevance and factuality metrics compared to both base models and those tuned individually, while maintaining competitive readability. Prompt design experiments highlight a trade-off between readability and factual accuracy. Our fine-tuned model demonstrates strong performance in relevance and factuality among open-source systems and rivals closed-source models such as GPT, providing an efficient and effective solution for domain-specific lay summarization.

Instruction tuning has underscored the significant potential of large language models (LLMs) in producing more human controllable and effective outputs in various domains. In this work, we focus on the data selection problem for task-specific instruction tuning of LLMs. Prevailing methods primarily rely on the crafted similarity metrics to select training data that aligns with the test data distribution. The goal is to minimize instruction tuning loss on the test data, ultimately improving performance on the target task. However, it has been widely observed that instruction tuning loss (i.e., cross-entropy loss for next token prediction) in LLMs often fails to exhibit a monotonic relationship with actual task performance. This misalignment undermines the effectiveness of current data selection methods for task-specific instruction tuning. To address this issue, we introduce ROSE, a novel Reward-Oriented inStruction data sElection method which leverages pairwise preference loss as a reward signal to optimize data selection for task-specific instruction tuning. Specifically, ROSE adapts an influence formulation to approximate the influence of training data points relative to a few-shot preference validation set to select the most task-related training data points. Experimental results show that by selecting just 5% of the training data using ROSE, our approach can achieve competitive results compared to fine-tuning with the full training dataset, and it surpasses other state-of-the-art data selection methods for task-specific instruction tuning. Our qualitative analysis further confirms the robust generalizability of our method across multiple benchmark datasets and diverse model architectures.